Upskale

What it does

Upskale enhances an image's resolution using Real-ESRGAN — a deep-learning super-resolution model that's become the de-facto open-source standard for natural-image upscaling. You drop in a picture, choose how much to enlarge it, and get back a sharper, larger version. The model is particularly good at preserving real-world textures (skin, fabric, foliage) without the over-smoothed plastic look that earlier bicubic or AI methods produced.

Two ways to upscale

• Scale factor mode — pick 2×, 3×, or 4× and Upskale enlarges the input by that exact factor. Good when you want predictable proportional output (e.g., turn a 1024×768 source into a 4096×3072 print master).
• Target resolution mode — pick 1080p, 2K, or 4K and Upskale aims for that final size, using a combination of AI upscaling and high-quality resampling to land exactly there.

Good use cases

• Restoring old family photos — scans of physical prints, low-resolution exports from old phones, screenshots from old social media accounts.
• Prepping small images for larger surfaces — a 500px logo that needs to live on a banner, a thumbnail that needs to fill a hero section, a product shot for an ecommerce listing.
• Cleaning up screenshots — slides, diagrams, or design references that came back blurry from a screen capture.
• Salvaging AI-generated art — pretty much every diffusion-model export benefits from a 2× pass through Real-ESRGAN.

Privacy

No account is required. Files you upload are processed on a server in Mumbai, written to disk only as long as the job is running, and deleted automatically — typically within an hour. The output is yours; we have no copy, no analytics on the contents, and no model fine-tuning happening behind the scenes.

Limits and rate limits

Upskale runs on a single consumer-grade GPU, so it's queue-based and gently rate-limited:

• Max input dimension: 4096 px on the longest side.
• Max upload size: 20 MB.
• Rate limit: 3 uploads per hour per IP. Generous for individual use, tight enough to keep the queue moving.
• Output format: PNG (lossless).

Under the hood

The model is Real-ESRGAN x4plus, running on PyTorch CUDA on an NVIDIA RTX-class GPU. The backend is a Node + Express server with SQLite for the job queue, and a Cloudflare tunnel exposing the local server to the internet. The frontend you see (when you click Launch) is a static page deployed on Vercel.